few-shot-learning

A PyTorch-based framework for few-shot learning and zero-shot image classification, designed for research and practical applications. This repository provides tools for training, evaluating, and experimenting with models that can generalize to new classes with limited labeled data.

Features

• Few-shot and Zero-shot Learning: Train models to recognize new classes with few or no labeled examples.
• Flexible Data Loading: Custom dataloaders for datasets like Flickr and NUS-WIDE.
• Word Embedding Integration: Uses word2vec for semantic class representations.
• Modular Model Design: Easily extend or modify the backbone and classifier.
• Custom Loss Functions: Includes semantic ranking loss for multi-label classification.

Project Structure

├── dataloaders/         # Data loading utilities
│   └── dataloader.py
├── models/              # Model architectures
│   └── model.py
├── utils/               # Utility functions and custom losses
│   ├── loss.py
│   └── util.py
├── preprocess.py        # Data preprocessing scripts
├── train.py             # Training script
├── LICENSE
└── README.md

Getting Started

Prerequisites

• Python 3.7+
• PyTorch >= 1.7
• torchvision
• numpy
• OpenCV
• PIL

Install dependencies:

pip install torch torchvision numpy opencv-python pillow

Data Preparation

• Download the NUS-WIDE or Flickr dataset.
• Place the dataset in your desired directory and update paths in preprocess.py and dataloaders/dataloader.py.
• Run preprocessing:

python preprocess.py

Training

Edit train.py to set your training configuration, then run:

python train.py --epochs 50 --saved_path ./checkpoints

Model

The main model is based on ResNet-101, with a custom classifier that projects image features into a semantic space using word2vec embeddings.

Customization

• To use a different backbone, modify models/model.py.
• To add new loss functions, edit utils/loss.py.

Citation

If you use this codebase in your research, please cite this repository.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contributing

Pull requests are welcome! For major changes, please open an issue first to discuss what you would like to change.

Contact

Maintainer: Md Hasan